Skip to content

concept.md

Latest commit

 

History

History
61 lines (40 loc) · 28.2 KB

concept.md

File metadata and controls

61 lines (40 loc) · 28.2 KB
Raw

title: Concept — Design Taste date: 2026-06-08 type: concept status: active updated: 2026-08-12 tags:

  • design
  • principles
  • distilled related:
  • "Design Taste Ledger"
  • "GitHub Trending Design Practice" sources:
  • "Design Taste Ledger"

Concept — Design Taste

Living distillation of what makes a landing/hero UI land for my audience (technical builders; AI-agent / dev-tool / PKM space). Built from repeated evidence in the Design Taste Ledger — not from vibes. The nightly continuous-learning loop updates this as patterns accumulate; every claim should cite the rep notes that support it.

[!note] Status: active Distilled from the Design Taste Ledger as patterns repeat across reps. Each principle cites ≥2 supporting rep notes.

Principles that keep proving out

  • Real/bespoke artifacts beat stock or faked ones. A genuine code snippet, a hand-built diagram, or honestly-labelled "illustrative" data lands far better than stock illustrations, div-fake "dashboards", or fabricated benchmarks. Proven repeatedly: bespoke animated SVG memory-graph (strongest of its day) and an illustrative-not-fabricated ranked panel 2026-06-08; a real from turbovec import Index code panel explicitly avoiding a div-fake 2026-06-09; a working instrument readout vs a screenshot 2026-06-10. Sharpened 2026-08-04 - the strongest real artifact is the product's own output, shown at full size. Two reps the same day landed on exactly this. A logging library made its own log lines the largest block on the page, so the product sold itself by showing what it writes. A realtime agent framework drew the gap between two turns as a waveform instead of quoting a latency figure, and the claim became visible with no benchmark number at all 2026-08-04. This is the sharpest form of the principle: a bespoke diagram is one step removed from the product, while the thing the tool emits is the product. It also explains earlier wins that looked like different moves - a real diff inside a terminal frame 2026-07-31 and a classifier verdict as the hero 2026-08-03 were both the output, not a picture of it. The test: could you have produced this artifact without running the tool? If yes, it is a depiction and something better is available. Qualified 2026-08-05 - the exception is when the real output hides the primitives. A computer-vision library landed by drawing its detection output as a drafting sheet with dimension lines rather than a real annotated frame, and the rep note records why: the schematic names the primitives better than a real frame does 2026-08-05. A skills repo the same day put an actual file card beside the claim so the abstraction had something concrete attached, and a state kernel drew its held goal and armed quota as gauges you could read 2026-08-05. So the rule is not always show the raw output. It is show the thing the reader needs to be able to name. When the output is self-explanatory, ship it raw. When the output is a picture whose meaning is in the labels, draw the labels. The failure both directions guards against is the same one: a visual the reader cannot interrogate.

  • Let the capability list be the layout (content-as-structure) instead of decorating around it. Rendering the substance as the page structure beats card-rows / logo-walls: a numbered "sources ledger" and a real install/execute/edit/test verb-loop both landed 2026-06-09; an agent roster as an editorial index and a prompt-excerpt reference document both landed again 2026-06-12. Four-plus clean wins — the most reliable move in the practice.

  • Make the hero visual diagram the value proposition (the mechanism), not decorate. The strongest mocks let one visual literally explain the claim: a memory-graph = the product 2026-06-08; a radar scope = "4 agents on one shared memory" and a loadout gauge = "skills on demand" 2026-06-10; a scan-result terminal = "catch danger" and a 4-node loop = "self-improve" 2026-06-11. A motif is tasteful when the instrument encodes the actual claim, and drifts to costume when it's a frame around ordinary copy. Now proven style-independent: the same "supporting artifact carries the hero, not the headline" finding held three days running across different style families — a terminal-dark live read-feed 2026-06-15, an editorial two-node dial(key) diagram 2026-06-16, and a literal HUD instrument-cluster of dials 2026-06-17 — so this is a general law of the practice, not a quirk of one aesthetic. Restated 2026-06-30/07-01 as "design in the artifact's native medium": show the mechanism in the medium it runs in — a routing console with provider lanes, a terminal walking create->evaluate->deploy, a Bauhaus token wall, an editorial for/against/abstain deliberation column, a frosted isolation panel of sealed agent cubes, an ASCII terminal multiplexer — six more landed across six different style families 2026-06-30 2026-07-01. Held again 2026-07-05/07-07: an annotated NL->DOM page schematic and a presence-radar sensor readout 2026-07-05, then a poster that lets type alone catalog a skills library, a radar that renders the sensor for camera-free sensing, and a Bauhaus shape-per-format system 2026-07-07 — the form chosen each time so the repo demonstrates its claim rather than asserting it. Reconfirmed 2026-07-12: a policy terminal visibly blocks danger/allows inspection, a calm queue embodies durable background work, and a strict Bauhaus composition demonstrates anti-slop rules — three more cases where the interface behaves like the system it represents 2026-07-12. Reconfirmed 2026-07-18: a streaming point field reveals reconstruction, a see→decide→refine composition teaches design practice, and a search→fetch→crawl→research terminal exposes the local-first workflow — three different styles, each credible because the visual demonstrates the mechanism 2026-07-18. Reconfirmed 2026-07-19: heterogeneous compute placement, a creator-controlled voice studio, and an explicit intent→plan→execute→verify harness each made trust legible by showing where work goes, who controls it, and how completion is proven 2026-07-19. Reconfirmed 2026-07-20: a quota route controller, a calm evidence/decision memory graph, and a small illustrative Python surface each made an abstract capability credible by exposing its mechanism without turning the page into game UI or documentation 2026-07-20. Reconfirmed 2026-07-21/23: an accountable agent roster and ontology diagram 2026-07-21, inspectable CAD/hardware-fit surfaces 2026-07-22, then visible coordination lanes and a live code-linked architecture map 2026-07-23 all landed by turning an abstract promise into a concrete operating surface. Reconfirmed 2026-07-25: a hybrid code-review pipeline that visibly separates deterministic rules from agent judgment, a measured multi-axis harness scorecard, and a tangible peer-to-peer mesh topology each landed across three unrelated style families (brutalist, data-viz, mono-zine) — in every case the visual exposed the mechanism rather than asserting the claim, and the credibility came from the mechanism being legible, not from the aesthetic 2026-07-25. Reconfirmed 2026-07-26: one unchanged call shape sitting above interchangeable provider targets made portability read as engineering leverage, and a compose->control->extend->deliver block system made self-hosted ownership feel structural — both landed by showing the mechanism operating rather than naming the benefit 2026-07-26.

  • Label faux/illustrative data honestly ("illustrative") — it then lands instead of misleading. Repeatedly an explicitly-labelled illustrative panel earned the "fast/ranked/live" claim without faking a benchmark: ranked output panel 2026-06-08; loadout gauge + radar readout both footed ILLUSTRATIVE 2026-06-10.

  • A stated constraint or a measured number outperforms a superlative — claim less, more precisely. The page gets more credible as the language gets less grand, because a number invites checking and an adjective invites doubt. Proven twice in one day: a performance kernel presented as a measured profile instead of superlative language landed by inviting verification rather than belief, and a coding tool made a memory budget the headline commitment so efficiency read as engineering intent rather than marketing 2026-07-29. The same move earned earlier wins — a cache-telemetry gauge honestly footed GAINS DEPEND ON WORKLOAD 2026-06-13 and a harness framed as measured axes rather than "best" 2026-07-25. Distinct from the honest-labelling principle above: that one is about not faking data, this one is about replacing the adjective with the figure. When reaching for "blazing", "powerful", or "seamless", state the constraint you actually hold instead. Extended 2026-08-07 - when you have no number, draw the shape and invent nothing. A context-reduction tool landed by showing two stacked bars with no percentage attached at all, and the rep note records why: the gap between the bars carries the argument and the chart stays credible 2026-08-07. A skills repo the same day made provenance the headline and put the actual directory beside the claim, so showing the directory proved it rather than restating it 2026-08-07. Both are the honest-data principle applied where the honest data is a relationship rather than a figure. A fabricated percentage would have been more specific and less true. The rule: if you have a number, lead with it; if you have a ratio you can show but not measure, show the ratio unlabelled; never manufacture precision to fill the gap. Widened 2026-07-30 — state the BOUNDARY, not just the number. The same axis extends past figures to limits: an unofficial training path landed because it labelled explicitly where the boundary was, so exploration read as rigorous rather than reckless, and frontier voice research landed by carrying one restrained orb and waveform instead of spectacle, keeping ambition legible without demo theatrics 2026-07-30. With the measured-profile kernel and the memory-budget headline 2026-07-29 that is four reps across two days on one finding: naming what you do not have, or where the work stops, buys more credibility than it costs. The instinct to hide the limit is exactly backwards — an unstated boundary reads as an unexamined one. Sharpened 2026-07-31 — restraint has to read as a CHOICE, not an absence. This is the failure mode lurking inside every principle above: hold back without showing why, and minimal reads as unfinished. A project manager landed by making restraint itself the product argument — three calm columns, no ceremony — so minimalism read as considered rather than incomplete, and frontier voice research landed with one restrained orb instead of spectacle, keeping ambition legible without demo theatrics 2026-07-31 2026-07-30. The test: could a viewer tell the difference between what you left out on purpose and what you never got to? If not, the restraint is not yet doing work. A skill-routing console the same day passed it by stating authorized use on the face of the dispatch view 2026-07-31. Fourth consecutive day 2026-08-02, and now the visual scales to the figure. A local-inference project charted the stated 4GB budget against 70B parameters with layer residency drawn low, using only numbers already in the repo description, and landed because the constraint was the argument with nothing invented to inflate it. A team-memory project on the same day named its four assets on one panel and let a governance line do the work adjectives usually attempt 2026-08-02. Read together with the memory-budget headline and the measured profile 2026-07-29, the stated boundary 2026-07-30, and restraint-as-a-choice 2026-07-31, the axis now has six reps across four days and one operational rule falls out of it: when the product claim is a number, that number gets the largest element on the page, and the chart around it may contain nothing the source did not state. The failure mode this rules out is the honest number wrapped in an invented visual, which spends the credibility the number just bought.

  • An instrument/HUD frame only lands when a real readout or code/command proof sits inside it — a bare frame around ordinary copy underdelivers. The aviation motif needs concrete content under it: a one-line model="openai:gpt-4o" # or "anthropic:…" swap-terminal made the reticle land and explicitly fixed the earlier bare-reticle miss 2026-06-13; an honestly-footed cache-telemetry gauge (GAINS DEPEND ON WORKLOAD) landed 2026-06-13; whereas a reticle whose "telemetry" was a slogan only mostly landed 2026-06-10. Frame + real proof = lands; frame alone = costume. Reconfirmed and sharpened 2026-07-27 — put the receipts under the claim: an instrument-sweep treatment of multi-community research landed specifically because the sources stayed visible beneath the synthesis, and a poster-scale capability announcement landed because it listed the mechanism explicitly instead of trading on scale alone 2026-07-27. The generalisation across both: the bolder the visual claim, the more literally the substantiation has to sit adjacent to it. Boldness buys attention; visible receipts buy belief. Held again 2026-07-28: a governance toolkit drew identity, policy, sandbox, and reliability as gates wrapped around one agent action so that enforcement read as architecture rather than compliance decoration, and a poster-scale book-to-skill claim landed by showing honest transformation states beneath the promise 2026-07-28 — the second consecutive day the same style family (poster/signal-red) landed on exactly this axis 2026-07-27. Governance specifically is the hardest case and now has two wins: a policy terminal that visibly blocked danger 2026-07-12 and this gate architecture 2026-07-28. Draw the enforcement, don't badge it. Extended 2026-07-29: an open collaboration stack shown as inspectable direction, execution, and record lanes made openness read as operational rather than ideological 2026-07-29 — the same move applied to a value rather than a mechanism, and it works for the same reason: you can look at the lanes.

  • Turn a tagline or claim into a literal staged sequence (pipeline / arc) — make the process the hero. When the pitch is really a sequence, rendering it as an explicit ordered arc lands: a 3-word tagline drawn as a numbered Learn->Build->Ship pipeline 2026-06-15 and a tagline staged as a literal Vibe->Decompose->Engineered arc 2026-06-19 both landed. Caveat — the stages must be real: a 5-stage run-ladder invented for a tagline-less repo only mostly landed because the metaphor carries more than the source proves 2026-06-20. Sharpened 2026-08-01: the arc is the fix for the catalogue. Twice in three days a project with a large inventory landed by rendering the path through it instead of the contents of it. A skills repo gave its frame->apply->confirm method equal billing with the headline and read as a working practice rather than a capability catalogue 2026-07-30, and a 24-lesson curriculum drew a ground->model->apply->extend geometric arc instead of a lesson index, which made the path feel deliberate and finishable 2026-08-01. Two different style families, editorial and Bauhaus, same move. This is the constructive form of the inventory anti-pattern below: when the honest content is a list, the fix is not a better list, it is the ordered route through it.

  • Editorial restraint (serif + generous whitespace + a single warm accent) builds credibility for methodology / thought-leadership pieces where a dark dev-tool look would undercut it. When the subject is a framework or a point of view rather than a tool, a calm editorial treatment reads as considered, not hyped: 18 personas as a newspaper deliberation council 2026-07-01 and an agentic methodology as a calm Georgia-serif manifesto with one burgundy italic word 2026-07-08 both landed. (Promoted from an open hypothesis once it landed twice across different repos.) Sharpened 2026-07-26 — the artifact should demonstrate its own thesis: a design-language repo presented as calm editorial reasoning with a single instruction sample landed precisely because it argued for judgment and exercised judgment at the same time 2026-07-26, echoing the learn->build->test manifesto that made practice feel consequential without pretending it was effortless 2026-07-24. For a methodology piece, restraint is not just a credibility signal — it is the proof. A loud treatment would refute the argument it is making. Held again 2026-07-30: a skills repo gave its frame->apply->confirm method equal billing with the headline claim and landed as a working practice rather than a capability catalogue 2026-07-30 — showing how you work outperforms listing what you can do, which is the same move one level up.

  • Warm-minimal lowers intimidation around technically dense infrastructure and learning — without hiding the mechanism. A cream or soft-neutral canvas lowers the emotional temperature while one concrete device/flow diagram still proves the claim: fully-local dictation as a calm waveform-to-text device 2026-06-28, a 100%-local meeting assistant as a soft transcription→summary flow 2026-07-05, and secure WireGuard+2FA as a warm three-node tailnet around a central trust check 2026-07-11. A fourth landed rep extends the pattern beyond security into systems learning: a warm choose→deconstruct→rebuild→explain workbench stayed approachable without diluting the technical premise 2026-07-17. Warmth is not decoration; it lowers the emotional temperature while concrete mechanism preserves rigor. Extended 2026-07-24: an offline grammar companion made privacy a quiet product behavior—not a badge—reconfirming that local/private tools land when calm presentation is paired with one concrete interaction 2026-07-24. Reconfirmed again 2026-07-27: a terminal file manager rendered calm and navigable through soft structure landed without surrendering its keyboard-first character 2026-07-27 — and a keyboard-first review tool landed the same way by proving itself with a real diff and comment inside the terminal frame, demonstrating in the medium reviewers already work in rather than selling them out of it 2026-07-31 — warmth lowered the barrier to a dense power-user tool rather than softening the tool itself, which is the distinction that keeps this principle from collapsing into "make it pretty".

  • Let ONE element carry the whole claim — singular focus beats accumulation. The strongest mocks repeatedly hang everything on a single artifact rather than assembling several: one unchanged call shape proving portability and one instruction sample proving a design language 2026-07-26, with the same device landing again six days later as one unchanged session shape sitting above the surfaces an embedded agent serves, so portability read as an engineering property rather than a slogan 2026-08-01; one local edit making privacy behavioral 2026-07-24; one central trust check anchoring a secure tailnet 2026-07-11; one concrete device/flow diagram carrying a fully-local assistant 2026-07-05; and three more the same day — gates around one agent action, one frosted listen→understand→respond stack, one poster claim 2026-07-28. This is the counterweight to the inventory anti-pattern below: a sample of many reads thin, but a single fully-realised instance reads complete. When tempted to add a second supporting element, sharpen the first instead. Sharpened again 2026-08-09 - make the RULE visible rather than the feature. Four reps across two days landed on the same move, and each time the single element was a constraint the system enforces. Single ownership was drawn as one lit cell across three node panels, and the rep note gives the reason: the rule you can see beats the rule explained in a paragraph 2026-08-08. A retrieval unit became one lit node with four typed edges, so retrieval by connection reads as a picture instead of a claim 2026-08-09. A tracing tool put its own question in the hero and ran the trace beside it, on the finding that one question answered beats any feature list 2026-08-09. And a multi-agent trading system showed the disagreement between seats rather than returns, where structure carried the credibility a fabricated number would have destroyed 2026-08-08. The distinction that took four reps to see: the decision principle says show what the system CONCLUDES, and this says show what it will NOT ALLOW. A conclusion is one output; a rule is every output at once, which is why it survives the reader asking what happens next. Sharpened 2026-08-03 - the one element should be the DECISION the system makes, not its speed or capacity. Three reps across two days converge on this. A document classifier made the verdict itself the hero and routed each file explicitly, and the rep note records why it worked: the decision is legible before any speed claim 2026-08-03. A local inference engine spent the entire page on one sentence with three backend names as the only proof, and restraint read as confidence because the claim stayed small enough to be true 2026-08-03. A multi-source reader put the access claim in the largest type and let a source rail carry the breadth underneath 2026-08-02. The common move is that the largest element on the page states what the system concludes or commits to, and everything else is evidence for it. This is the answer to the question the singular-focus principle leaves open: once you accept that one element carries the page, the decision is the thing that element should be. It also explains why generic stat-tiles keep losing below - a throughput number is an attribute, and an attribute cannot be the hero because nothing else on the page can support it. Sharpened 2026-08-11 - make the singled-out element the odd state among peers, not a lone hero. Four consecutive days landed the same move across four unrelated style families: one lit cell across three node panels for single ownership 2026-08-08, one lit node with four typed edges for retrieval by connection 2026-08-09, a needs-you row given equal standing beside running and idle so the paused row is where the eye lands 2026-08-10, and five agent lanes at different depths with one blocked, which made parallel read as a scheduling problem rather than a feature word 2026-08-11. The peers are what make it work. A single hero element asserts the claim; a row of peers with one that differs lets the reader derive the rule from the contrast, and a rule the reader derived is one they now believe. The test: remove the peers and see whether the highlighted element still says anything.

  • Pick ONE word out of the project's own description and build the layout to defend it. Four landed reps across two days used the same move, and in two of them the rep note names the word outright. A skills repo turned on the word public, showing what changes once the file is readable rather than listing what the skills do 2026-08-11. An agent runner argued its category name, laying out five lanes at different depths with one blocked so parallel read as a scheduling problem instead of a feature word 2026-08-11. A retrieval engine was built around the word layer, two stacked bars comparing what fills the context window when retrieval is the whole story against when tool results and prior steps also land there, which is what makes layer the right noun 2026-08-12. A workspace product took at-linked and ran one identifier across email, a call and a task so the memory row at the bottom is where they resolve to one record 2026-08-12. The counter-move is illustrating the whole sentence, which is how a hero ends up as a feature list with better type: the macro tagline names seven products, and rendering all seven would have produced a list, while one identifier surviving three surfaces is an argument for calling it a single workspace. The test: read the description, find the word doing the most work and least defended, and make the page prove that word. This sits underneath the singular-focus principle rather than beside it. Singular focus says one element carries the page and the decision principle says that element should be what the system concludes. This says where to look for the claim in the first place, which is the vocabulary the project already chose.

Anti-patterns (repeated misses)

  • A small sample of an inventory (a few-row rack, an N-node grid) reads thin/toy — show one produced result/state instead. When the artifact depicts a representative slice of a large set, it underdelivers because the slice looks like the whole: a four-row "skills rack" mostly landed but the sample "risks looking thin" 2026-06-18; an eight-pod cluster grid mostly landed but "edges toward toy" 2026-06-17. The reliable fix is to render a concrete result the product produces rather than a roster of what it contains — a knowledge-graph readout 2026-06-18 and a read->edit->test diff transcript 2026-06-18 both landed the same day the rack only mostly did, and an instrument readout 2026-06-17 beat the pod grid of its day. Show state/result, not inventory. Corollary — uniform symmetric / parallel N-card layouts trip the same wire: equal-weight cards in a tidy row read as a template even when the copy is good — a Build·Ship·Iterate three-card triad mostly landed but "symmetry risks looking template" 2026-06-19, echoing the four-row rack of the day before 2026-06-18. Break the symmetry (vary weight/size, or replace one card with a real produced result) instead of shipping a clean parallel grid. Fix found 2026-08-06 - swap the card grid for a ruled ledger. The same four-asset memory inventory that landed as a frosted panel on 2026-08-02 was retried as a ruled ledger with a governed status column, and the rep note records why it worked better: a ledger implies accountability where a card grid does not 2026-08-06. That is the same content in the same quantity, so the change is purely the container. A grid says here are four things. A ledger says here are four things and someone is answerable for each. When the inventory is genuinely the point and cannot be replaced by a produced result, ruling it and adding a status column is the move that stops it reading as a template.

  • Generic stat-tile dashboards / borrowed product-template heroes underperform. A stat-tiles + sparkline block "reads template" and a telemetry strip that's really a slogan dressed as a stat both only mostly landed — the weakest moves of their days: telemetry "It ships" cell 2026-06-10; agentsview stat-tiles "the most generic move in the deck" 2026-06-11; and re-skinning that same dashboard into a HUD "session cockpit" still left it the weakest shape of its day 2026-06-13 — proof that HUD styling lifts but does not rescue a stats-in-tiles hero. Same axis as the win above: bespoke-instrument beats generic-dashboard.

  • A familiar/centered hero layout underperforms even when the execution is clean. Whenever the hero falls back to a generic centered headline+CTA, the mock only mostly lands and is the weakest of its day, regardless of polish: a centered manifesto "leans on the headline" 2026-06-09; a big centered launch serif was the day's fragile one 2026-06-12; a clean single-accent dark hero was "a touch familiar" and the weakest of three 2026-06-14; a centered CUA hero anchored by 3 capability cards only mostly landed because "centered = the safe choice" — an actual sandbox-screen visual would have raised it 2026-06-15. The fix is the same as the wins above — let a bespoke, product-specific visual carry the hero instead of a centered template.

  • A visual metaphor that claims more than the repo actually proves only mostly lands — scale the ambition of the mock to the available evidence. When the hero motif outruns what the source substantiates it reads as a disclaimer-propped stretch: a static waveform player leaned on its illustrative footnote 2026-06-16 and an invented 5-stage harness ladder carried more than the source proves 2026-06-20 — both only mostly landed. Same axis as the honest-data principle above: pick a motif the repo can actually back, or scale it to what is provable.

Open hypotheses

  • Editorial style may build more credibility for thought-leadership → promoted to a proven principle 2026-07-08 (landed twice: 2026-07-01, 2026-07-08).
  • HUD style is high-variance — strong when restrained, costume-y when overdone.